[dxvk] Make use of VK_AMD_rasterization_order

May slightly improve GPU performance in some scenarios.

src/dxvk/dxvk_compute.cpp

@@ -1,20 +1,22 @@
 #include "dxvk_compute.h"
+#include "dxvk_device.h"
 
 namespace dxvk {
   
   DxvkComputePipeline::DxvkComputePipeline(
-    const Rc<vk::DeviceFn>&       vkd,
+    const DxvkDevice*             device,
     const Rc<DxvkPipelineCache>&  cache,
     const Rc<DxvkShader>&         cs)
-  : m_vkd(vkd), m_cache(cache) {
+  : m_device(device), m_vkd(device->vkd()),
+    m_cache(cache) {
     DxvkDescriptorSlotMapping slotMapping;
     cs->defineResourceSlots(slotMapping);
     
-    m_layout = new DxvkBindingLayout(vkd,
+    m_layout = new DxvkBindingLayout(m_vkd,
       slotMapping.bindingCount(),
       slotMapping.bindingInfos());
     
-    m_cs = cs->createShaderModule(vkd, slotMapping);
+    m_cs = cs->createShaderModule(m_vkd, slotMapping);
     
     this->compilePipeline();
   }

src/dxvk/dxvk_compute.h

@@ -7,6 +7,8 @@
 
 namespace dxvk {
   
+  class DxvkDevice;
+  
   /**
    * \brief Compute pipeline
    * 
@@ -20,7 +22,7 @@ namespace dxvk {
   public:
     
     DxvkComputePipeline(
-      const Rc<vk::DeviceFn>&       vkd,
+      const DxvkDevice*             device,
       const Rc<DxvkPipelineCache>&  cache,
       const Rc<DxvkShader>&         cs);
     ~DxvkComputePipeline();
@@ -47,7 +49,9 @@ namespace dxvk {
     
   private:
     
-    Rc<vk::DeviceFn>      m_vkd;
+    const DxvkDevice* const m_device;
+    const Rc<vk::DeviceFn>  m_vkd;
+    
     Rc<DxvkPipelineCache> m_cache;
     Rc<DxvkBindingLayout> m_layout;
     Rc<DxvkShaderModule>  m_cs;

src/dxvk/dxvk_device.cpp

@@ -15,7 +15,7 @@ namespace dxvk {
     m_memory          (new DxvkMemoryAllocator(adapter, vkd)),
     m_renderPassPool  (new DxvkRenderPassPool (vkd)),
     m_pipelineCache   (new DxvkPipelineCache  (vkd)),
-    m_pipelineManager (new DxvkPipelineManager(vkd)),
+    m_pipelineManager (new DxvkPipelineManager(this)),
     m_submissionQueue (this) {
     m_options.adjustAppOptions(env::getExeName());
     m_options.adjustDeviceOptions(m_adapter);

src/dxvk/dxvk_graphics.cpp

@@ -1,5 +1,6 @@
 #include <cstring>
 
+#include "dxvk_device.h"
 #include "dxvk_graphics.h"
 
 namespace dxvk {
@@ -33,14 +34,15 @@ namespace dxvk {
   
   
   DxvkGraphicsPipeline::DxvkGraphicsPipeline(
-      const Rc<vk::DeviceFn>&       vkd,
+    const DxvkDevice*             device,
-      const Rc<DxvkPipelineCache>&  cache,
+    const Rc<DxvkPipelineCache>&  cache,
-      const Rc<DxvkShader>&         vs,
+    const Rc<DxvkShader>&         vs,
-      const Rc<DxvkShader>&         tcs,
+    const Rc<DxvkShader>&         tcs,
-      const Rc<DxvkShader>&         tes,
+    const Rc<DxvkShader>&         tes,
-      const Rc<DxvkShader>&         gs,
+    const Rc<DxvkShader>&         gs,
-      const Rc<DxvkShader>&         fs)
+    const Rc<DxvkShader>&         fs)
-  : m_vkd(vkd), m_cache(cache) {
+  : m_device(device), m_vkd(device->vkd()),
+    m_cache(cache) {
     DxvkDescriptorSlotMapping slotMapping;
     if (vs  != nullptr) vs ->defineResourceSlots(slotMapping);
     if (tcs != nullptr) tcs->defineResourceSlots(slotMapping);
@@ -48,15 +50,15 @@ namespace dxvk {
     if (gs  != nullptr) gs ->defineResourceSlots(slotMapping);
     if (fs  != nullptr) fs ->defineResourceSlots(slotMapping);
     
-    m_layout = new DxvkBindingLayout(vkd,
+    m_layout = new DxvkBindingLayout(m_vkd,
       slotMapping.bindingCount(),
       slotMapping.bindingInfos());
     
-    if (vs  != nullptr) m_vs  = vs ->createShaderModule(vkd, slotMapping);
+    if (vs  != nullptr) m_vs  = vs ->createShaderModule(m_vkd, slotMapping);
-    if (tcs != nullptr) m_tcs = tcs->createShaderModule(vkd, slotMapping);
+    if (tcs != nullptr) m_tcs = tcs->createShaderModule(m_vkd, slotMapping);
-    if (tes != nullptr) m_tes = tes->createShaderModule(vkd, slotMapping);
+    if (tes != nullptr) m_tes = tes->createShaderModule(m_vkd, slotMapping);
-    if (gs  != nullptr) m_gs  = gs ->createShaderModule(vkd, slotMapping);
+    if (gs  != nullptr) m_gs  = gs ->createShaderModule(m_vkd, slotMapping);
-    if (fs  != nullptr) m_fs  = fs ->createShaderModule(vkd, slotMapping);
+    if (fs  != nullptr) m_fs  = fs ->createShaderModule(m_vkd, slotMapping);
     
     m_vsIn  = vs != nullptr ? vs->interfaceSlots().inputSlots  : 0;
     m_fsOut = fs != nullptr ? fs->interfaceSlots().outputSlots : 0;
@@ -146,9 +148,14 @@ namespace dxvk {
     vpInfo.scissorCount           = state.rsViewportCount;
     vpInfo.pScissors              = nullptr;
     
+    VkPipelineRasterizationStateRasterizationOrderAMD rsOrder;
+    rsOrder.sType                 = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_RASTERIZATION_ORDER_AMD;
+    rsOrder.pNext                 = nullptr;
+    rsOrder.rasterizationOrder    = this->pickRasterizationOrder(state);
+    
     VkPipelineRasterizationStateCreateInfo rsInfo;
     rsInfo.sType                  = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO;
-    rsInfo.pNext                  = nullptr;
+    rsInfo.pNext                  = m_device->extensions().amdRasterizationOrder.enabled() ? &rsOrder : rsOrder.pNext;
     rsInfo.flags                  = 0;
     rsInfo.depthClampEnable       = state.rsEnableDepthClamp;
     rsInfo.rasterizerDiscardEnable= state.rsEnableDiscard;
@@ -260,4 +267,29 @@ namespace dxvk {
     return true;
   }
   
+  
+  VkRasterizationOrderAMD DxvkGraphicsPipeline::pickRasterizationOrder(
+    const DxvkGraphicsPipelineStateInfo& state) const {
+    // If blending is not enabled, we can enable out-of-order
+    // rasterization for certain depth-compare modes.
+    bool blendingEnabled = false;
+    
+    for (uint32_t i = 0; i < MaxNumRenderTargets; i++) {
+      if (m_fsOut & (1u << i))
+        blendingEnabled |= state.omBlendAttachments[i].blendEnable;
+    }
+    
+    if (!blendingEnabled) {
+      if (m_device->hasOption(DxvkOption::AssumeNoZfight))
+        return VK_RASTERIZATION_ORDER_RELAXED_AMD;
+      
+      if (state.dsDepthCompareOp == VK_COMPARE_OP_NEVER
+       || state.dsDepthCompareOp == VK_COMPARE_OP_LESS
+       || state.dsDepthCompareOp == VK_COMPARE_OP_GREATER)
+        return VK_RASTERIZATION_ORDER_RELAXED_AMD;
+    }
+    
+    return VK_RASTERIZATION_ORDER_STRICT_AMD;
+  }
+  
 }

src/dxvk/dxvk_graphics.h

@@ -12,6 +12,8 @@
 
 namespace dxvk {
   
+  class DxvkDevice;
+  
   /**
    * \brief Graphics pipeline state info
    * 
@@ -72,7 +74,7 @@ namespace dxvk {
     VkBool32                            omEnableLogicOp;
     VkLogicOp                           omLogicOp;
     VkRenderPass                        omRenderPass;
-    VkPipelineColorBlendAttachmentState omBlendAttachments[DxvkLimits::MaxNumRenderTargets];
+    VkPipelineColorBlendAttachmentState omBlendAttachments[MaxNumRenderTargets];
   };
   
   
@@ -88,7 +90,7 @@ namespace dxvk {
   public:
     
     DxvkGraphicsPipeline(
-      const Rc<vk::DeviceFn>&       vkd,
+      const DxvkDevice*             device,
       const Rc<DxvkPipelineCache>&  cache,
       const Rc<DxvkShader>&         vs,
       const Rc<DxvkShader>&         tcs,
@@ -127,7 +129,9 @@ namespace dxvk {
       VkPipeline                    pipeline;
     };
     
-    Rc<vk::DeviceFn>      m_vkd;
+    const DxvkDevice* const m_device;
+    const Rc<vk::DeviceFn>  m_vkd;
+    
     Rc<DxvkPipelineCache> m_cache;
     Rc<DxvkBindingLayout> m_layout;
     
@@ -154,6 +158,9 @@ namespace dxvk {
     bool validatePipelineState(
       const DxvkGraphicsPipelineStateInfo& state) const;
     
+    VkRasterizationOrderAMD pickRasterizationOrder(
+      const DxvkGraphicsPipelineStateInfo& state) const;
+    
   };
   
 }

src/dxvk/dxvk_options.cpp

@@ -5,7 +5,7 @@
 namespace dxvk {
   
   const static std::unordered_map<std::string, DxvkOptionSet> g_appOptions = {{
-    
+    { "NieRAutomata.exe", DxvkOptionSet(DxvkOption::AssumeNoZfight) },
   }};
   
   

src/dxvk/dxvk_pipemanager.cpp

@@ -35,8 +35,8 @@ namespace dxvk {
   }
   
   
-  DxvkPipelineManager::DxvkPipelineManager(const Rc<vk::DeviceFn>& vkd)
+  DxvkPipelineManager::DxvkPipelineManager(const DxvkDevice* device)
-  : m_vkd(vkd) {
+  : m_device(device) {
     
   }
   
@@ -62,7 +62,7 @@ namespace dxvk {
       return pair->second;
     
     const Rc<DxvkComputePipeline> pipeline
-      = new DxvkComputePipeline(m_vkd, cache, cs);
+      = new DxvkComputePipeline(m_device, cache, cs);
     m_computePipelines.insert(std::make_pair(key, pipeline));
     return pipeline;
   }
@@ -92,7 +92,7 @@ namespace dxvk {
       return pair->second;
     
     const Rc<DxvkGraphicsPipeline> pipeline
-      = new DxvkGraphicsPipeline(m_vkd, cache, vs, tcs, tes, gs, fs);
+      = new DxvkGraphicsPipeline(m_device, cache, vs, tcs, tes, gs, fs);
     m_graphicsPipelines.insert(std::make_pair(key, pipeline));
     return pipeline;
   }

src/dxvk/dxvk_pipemanager.h

@@ -59,8 +59,7 @@ namespace dxvk {
     
   public:
     
-    DxvkPipelineManager(
+    DxvkPipelineManager(const DxvkDevice* device);
-      const Rc<vk::DeviceFn>&   vkd);
     ~DxvkPipelineManager();
     
     /**
@@ -99,7 +98,7 @@ namespace dxvk {
     
   private:
     
-    const Rc<vk::DeviceFn> m_vkd;
+    const DxvkDevice* m_device;
     
     std::mutex m_mutex;